inference/research
verifiedmedium · 0.75updated 2026-10-15

Argonne MI250 llama.cpp Qwen2-72B b1

4.6 tok/s aggregate (batch*(in+out)/lat)

Fields

id
benchmarks/argonne-mi250-llamacpp-qwen2-72b-fp16-b1-128io-tps
Metric
tps_aggregate
Value
4.6
Unit
tok/s aggregate (batch*(in+out)/lat)
Model
Qwen/Qwen2-72B
Methodology
THIRD-PARTY MEASURED (Argonne LLM-Inference-Bench arXiv:2411.00136, All_results.csv). llama.cpp, FP16/BF16, 128in/128out, AGGREGATE over batch. 1 tile bs1 lat 55.84s. 72B on one tile barely fits. CSV 'AMD MI250 GPU' = Polaris MI250.
Measured by
third_party
Reproducible
yes

Cross-references